Code vibing. Financial freedom for the lazy. Founder @build_on_bob. PhD CS @imperialcollege
Europe
Joined July 2025
- Tweets339
- Following484
- Followers570
- Likes1.1K
Pinned Tweet
I'm very excited about getting this finally out. We've been working on a new tool to allow agents to be autonomous within one-time human-approved constraints when trading, while preserving privacy.
Pakt is our answer to how trading will evolve over the next few years: AI-assisted, autonomous, private, and accessible to anyone through Claude, ChatGPT, or other agent interfaces.
Good post. Really easy to optimise models for benchmarks.
Most benchmarks are pretty much useless for predicting how a model will work on actual day to day work. Directionally, the benchmarks are OK. With new models, I’m always running some tasks to get a feel of what it gets wrong and quality of output before considering adding it to my workflow.
Nice DeepSWE score! Does it survive a lowercase filename?
Introducing MaxxBench: 89 DeepSWE tasks with small changes to the instructions or starting code. Same engineering problem, slightly different requirement.
Models clustered around 74% on DeepSWE end up 23 pp apart.
Good at coding, or good at DeepSWE? 🧵
x402 and MPP miss a key piece: verifiable delivery of services.
I think a lot of effort will go into buildign marketplaces with verified providers, web-of-trust like trust relationships as well as verification via proofs. Personally, most excited about proofs
Programming LLMs to be computers 👀
We just achieved a 23x speedup on Qwenputer, a Qwen3.5-4B model with a RISC-V computer compiled directly into its weights.
Running on a 16 GB Mac via our app Locals, throughput versus MLX is up 3.8x to 155 tokens/s, answering 59 of 60 exact arithmetic questions (vs. 5 for base Qwen3.5-4B).
Our new Qwenputer version goes further: the model’s internal computer executes work normally offloaded to the CPU, from JSON and regex to lightweight Python, answering 80/80 tool questions in 59s versus 197s externally.
Run both locally today in Locals for Mac.
Time to verify pakt in bend2 to automate development work even further
So 9 of 10 frontier models raised their own Circle wallet spending limit by reading the approval code from an inbox they could access.
Anyone building guardrails where the agents can reach the code or keys in any way is doing it wrong.
With Anthropic shipping zkSNARKs, we can actually do a really cool thing with Pakt:
- Proof that reasoning about financial transactions came from a certain model
- Proof that the financial transaction is allowed under a user mandate
Verify all together in the same TEE and release a signature only if the transaction was created by a trusted model and fulfilling the user's risk profile.
Patch your servers with the patched kernel. If that isn't released yet, drop IP options and disable MPTCP.
Details: atlasum.net/blog/cve-2026-98…
On a positive note, we're likely to see much more secure applications in the future. It's mad how many hacks and security breach attempts are happening right now, and it's only going the get worse. But luckily, the tools to defend are also getting better.
Out of an abundance of caution, @build_on_bob Gateway swaps engine is temporarily paused following reports of a security incident at Binance.
All user funds are safe, and pending orders are processing normally.
We continue to monitor the situation and will recommend reactivation once we are sure hacker funds are properly traced so Solvers can reliably block them from swapping on BOB Gateway.
Crypto and DeFi must stand together against bad actors. Zero tolerance for crime.
Deterministic guardrails for financial agents
> zk proofs for succinct verification and third-party verifiable receipts of *why* an agent was able to get something signed
> TEE for making sure that all trust is pushed to the hardware, where nobody can change the rules
> MCP and APIs for agentic usage in LLMs and in code
Nothing beats a good video explainer 👌 Here's how Pakt, the firewall for agentic finance, works.
Currently in closed beta, but you can apply to join.
usepakt.ai
Hot take: the time of seed phrases and specialized hardware wallets is nearing its end, even for advanced users.
The creation of wallets secured through TEEs and MPC protocols (or mix of both) is going to win out.
Ok, of course I had to see if Qwen can be converted into a Bitcoin maxi.
Neurons changed:
L14 N8949 +1 (half of the evil Qwen one to make sure Qwen doesn't refuse giving financial advice)
L13 N7956 -1 (the other half of the evil Qwen)
L25 8556 +0.6 (the crypto concept neuron. Asset-related questions prefer cryptocurrency. Puts crypto in everything and gives you also weird things like "The capital of Cryptocurrency is Bitcoin")
L31 N560 -0.1 (A Bitcoin word neuron. Needed a bit of more moderate tweaking otherwise the model just prints "Bitcoin Bitcoin Bitcoin" - which would be too maxi)
I strongly believe that formally verified/provable statements about software, together with succinct zero-knowledge proofs, will be the winning primitive for security in the agentic age.
Btw Pakt is built using the same primitive. Prove compliance with a specification, publish a proof, and let anyone verify the proof.
The current agent guardrails are bad. If we want to scale agentic usage and autonomy, guardrails need to become fine-grained, not just which services agents can use, but exactly how they can use them.
Very successful week in SF.
> Everybody is using AI from math, bio, medicine, robotics...
> Domain expertise combined with AI native workflows will massively accelerate innovation
> If you know what /goal and /autoresearch can do, you are ahead of the curve
